Rubin platform, Vera CPU, GR00T robots, and Alpamayo L4 autonomy — TFN

NVIDIA arrived on the ongoing CES 2026 as a really completely different firm from the one individuals as soon as knew — a producer of graphics playing cards.

Nevertheless, over the previous few years, it has quietly overhauled itself round synthetic intelligence, high-performance computing, and large-scale techniques that energy every part from knowledge centres to automobiles and robots.

CES, which is underway now in Las Vegas, gave the corporate a worldwide stage to point out how far that transformation has gone.

As a substitute of specializing in a single product line, the California-based tech firm held a broad set of bulletins spanning AI supercomputers, networking, storage, open fashions, autonomous driving, robotics, gaming, and industrial software program.

Many of those updates should not consumer-facing merchandise however constructing blocks meant for builders, enterprises, and companions shaping the subsequent technology of AI techniques.

NVIDIA founder and CEO Jensen Huang took the stage on the Fontainebleau Las Vegas to open CES 2026, declaring that AI is scaling into each area and each system.

Try the roundup of what NVIDIA revealed beneath

Rubin platform constructed round “excessive codesign”

NVIDIA launched the Rubin platform, constructed round “excessive codesign,” that means the corporate designed a number of core elements collectively quite than treating them as separate components.

Rubin combines a Vera CPU, Rubin GPU, NVLink 6 Swap, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Swap.

NVIDIA says Rubin targets massive workloads comparable to mixture-of-experts fashions, agentic techniques, and long-context reasoning, with claims of as much as 10x decrease per-token inference value and fewer GPUs required for some coaching workloads than Blackwell.

DGX SuperPOD: the “default” design for Rubin-scale deployments

NVIDIA has launched the DGX SuperPOD as a key setup for Rubin-based techniques utilized in companies and analysis.

These DGX Rubin techniques combine computing, networking, and software program to facilitate coaching, inference, and long-context reasoning, claiming to be “as much as 10x” cheaper in token prices in comparison with Blackwell.

DGX Spark + DGX Station: deskside techniques for giant fashions

NVIDIA additionally targeted on smaller, native techniques: DGX Spark and DGX Station are “deskside AI supercomputers” designed to run open-source and frontier fashions domestically, then scale workloads to the cloud as wanted.

The corporate mentioned DGX Spark can deal with 100B-parameter fashions, and DGX Station can deal with as much as 1T-parameter fashions.

BlueField-4 + “context reminiscence storage”

NVIDIA introduced a brand new platform known as Inference Context Reminiscence Storage, powered by BlueField-4. This platform addresses the problem of dealing with massive context knowledge generated by long-context, multi-agent AI fashions, which don’t match properly in conventional GPU reminiscence.

The corporate claims that this new platform improves reminiscence capability and permits sharing of context throughout massive clusters, providing as much as 5 occasions extra tokens per second and 5 occasions higher energy effectivity than conventional storage strategies.

BlueField-4 is predicted to be accessible within the second half of 2026, with partnerships together with Dell, HPE, IBM, Nutanix, Pure Storage, Supermicro, VAST Information, WEKA, and others.

Enterprise AI Manufacturing unit replace

NVIDIA has added new options to its “Enterprise AI Manufacturing unit validated design” to boost safety and pace up infrastructure.

The mixing now consists of platforms like Armis, Examine Level, F5, Fortinet, Palo Alto Networks, Rafay, Pink Hat OpenShift, Spectro Cloud (PaletteAI), and Pattern Micro.

These additions present advantages comparable to telemetry through NVIDIA DOCA Argus and improved workload isolation in Kubernetes environments.

Nemotron, Cosmos, Alpamayo, GR00T, Clara

NVIDIA introduced a spread of recent open sources, together with fashions, datasets, and coaching instruments for varied fields. They showcased a number of households of expertise: Nemotron (AI brokers), Cosmos (bodily AI), Alpamayo (self-driving autos), Isaac GR00T (robotics), and Clara (biomedical).

In addition they shared spectacular figures for his or her open knowledge contributions, together with 10 trillion language tokens, 500,000 robotics paths, 455,000 protein buildings, and 100TB of car sensor knowledge.

Alpamayo for autonomous driving

NVIDIA launched the Alpamayo household for autonomous driving, which incorporates AI fashions, simulation instruments, and datasets specializing in uncommon and sophisticated driving situations.

The important thing elements are Alpamayo 1, AlpaSim, and “Bodily AI Open Datasets.” Alpamayo is a reasoning mannequin designed to develop autonomous autos. Early collaborators on this challenge embody JLR, Lucid, Uber, and Berkeley DeepDrive, all working in direction of “stage 4” autonomy.

DRIVE Hyperion ecosystem

NVIDIA is increasing its DRIVE Hyperion ecosystem and has introduced partnerships with main suppliers and sensor firms, together with Aeva, Bosch, and Sony.

The DRIVE Hyperion is a ready-to-use structure that mixes computing and sensors, that includes two DRIVE AGX Thor chips that ship over 2,000 TFLOPS for superior sensor processing and real-time duties.

NVIDIA DRIVE AV software program goes into manufacturing

NVIDIA mentioned its DRIVE AV software program will debut within the all-new Mercedes-Benz CLA, beginning with an “enhanced stage 2” driver-assistance system anticipated on U.S. roads by the top of this yr. The CLA can also be described as the primary Mercedes mannequin to make use of the MB platform.OS platform.

The corporate described its “dual-stack” strategy: an end-to-end AI driving stack paired with a classical security stack constructed on NVIDIA Halos for redundancy and guardrails.

It additionally listed capabilities like point-to-point city navigation, proactive collision avoidance, and automatic parking, and pointed to a cloud-to-car pipeline utilizing DGX for coaching, Omniverse/Cosmos for simulation, and DRIVE AGX + Hyperion in-vehicle compute.

Siemens partnership enlargement

Siemens and NVIDIA introduced an expanded partnership to deliver AI deeper into industrial workflows, together with bodily AI.

The discharge states that NVIDIA will present AI infrastructure, simulation libraries, fashions, frameworks, and blueprints, whereas Siemens will commit “lots of” of commercial AI consultants, in addition to {hardware}/software program capabilities.

Either side framed this round digital twins, sooner product improvement, and real-time manufacturing adaptation.

RTX AI video technology on PC

NVIDIA showcased its creator instruments for native AI video technology utilizing LTX-2, claiming they’ll produce “as much as 20 seconds of 4K video” with built-in audio and multi-keyframe help.

In addition they collaborated with ComfyUI to enhance GPU efficiency by 40% and added help for NVFP4 and NVFP8 codecs, which assist cut back pace and VRAM utilization on RTX 50 Sequence graphics playing cards.

Gaming show and rendering updates

NVIDIA’s gaming replace highlighted new rendering and show options. It launched DLSS 4.5, which features a “6X” mode that may produce as much as 5 extra frames per rendered body, with availability anticipated in spring.

Moreover, they introduced that G-SYNC Pulsar screens at the moment are accessible, that includes over 1,000Hz efficient movement readability and G-SYNC Ambient Adaptive Know-how.